Oliver Ratmann

Making inference of generalised contact matrices feasible

Understanding and quantifying changing patterns in human contact behaviour is an important component for modelling of infectious disease dynamics, but also in sociology, and complexity science. A particular open challenge involves how social contact patterns can be estimated beyond the usual 2D age-age dimensions to capture important further variation by socio-economic status, household size, multi-dimensional indicators of deprivation, or race and ethnicity. I will introduce Bayesian models for this task, that are similar in spirit to geospatial models but also satisfy inherent flow constraints. Interestingly, these constraints enable learning beyond the directly observed data by exploiting symmetries, sum-to-zero conditions and tabular constrains emerging from fixed marginals. I will also cover techniques that allow us to satisfy these constraints at no cost, opening up the use of state-of-the-art variational inference or HMC samplers, and neural architecture approximations. Time permitting, I will present first applications to social contact survey data. This is joint work with Shozen Dan, Zhi Ling and Swapnil Mishra on behalf of the Machine Learning & Global Health network.